@audio/decode-mpc

Decode Musepack (MPC) to PCM float samples. libmpcdec (mirrored at evpobr/libmpc) compiled to a single-file WASM ES module — no side files, loads from any CDN, Node, workers and AudioWorklets.

import decode, { decoder } from '@audio/decode-mpc'

let { channelData, sampleRate } = await decode(mpcBytes)

let dec = await decoder()
let head = dec.decode(chunk1)   // synchronous; returns whatever the bytes seen so far allow
let tail = dec.flush()
dec.free()

Decodes both stream generations: SV7 (MP+ magic, 1999–2006 era encoders) and SV8 (MPCK magic, current). The demuxer (container framing, APEv2/ID3v1 tags, seek tables) runs inside libmpcdec itself; this package only feeds it bytes. Mono or stereo, whatever the stream declares. Built with MPC_FIXED_POINT off, so libmpcdec hands back native float samples — no fixed-point requantization step.

API

decode(src: Uint8Array | ArrayBuffer): Promise<AudioData>

Decode a complete stream. Throws if the input never parses as Musepack (no MP+/MPCK magic, or a header that still won't parse after a generous budget).

decoder(): Promise<MpcDecoder>

Streaming decoder: decode(chunk) returns the samples of every complete frame the data seen so far allows, flush() decodes anything a safety margin was withholding at the true end of stream, free() releases WASM memory. errors counts frames abandoned after a bitstream desync (see Streaming below — this stops the decoder, it does not resync and skip).

AudioData

{ channelData: Float32Array[], sampleRate: number }

Streaming

decode(chunk) feeds bytes into a growable buffer inside WASM and retries mpc_demux_init() on every call until the header parses — libmpcdec's mpc_reader callback interface has no "not enough data yet" signal distinct from "end of stream," and a frame decode that runs out of buffered bytes mid-frame leaves the bitstream position permanently desynced (no retry). So frames are only decoded once at least 4352 bytes (MAX_FRAME_SIZE, libmpcdec's own worst-case compressed frame size) are buffered ahead of the reader's cursor — safe for any chunk size, including 1 byte at a time.

What that means in practice differs by version:

• SV7 streams genuinely progressively. Its header is a fixed ~30 bytes with no forward references, so decoder() starts emitting frames as soon as the first chunk carries that much.
• SV8 streams keep their seek table near the end of the file — mpcenc can only patch the header's pointer to it once encoding finishes and the final byte offset is known — and libmpcdec reads that table during mpc_demux_init(). In practice an SV8 decoder() buffers (close to) the whole stream before the first frame comes out. decode(chunk) never blocks or throws while this is happening; it just returns nothing until the header resolves.

Either way the input buffer is never discarded (the SV8 seek-table read can jump backward to any earlier offset), so streaming does not save memory for SV8 — only latency-to-first-decode for SV7.

Measured

• 50,773 bytes WASM (libmpcdec 7d04739, -Oz -flto, single-file ES module).
• ~86–89 dB SNR vs ffmpeg -i x.mpc -f f32le - on SV8 fixtures generated with mpcenc; 57–58 dB on an independent SV7 sample from the FFmpeg FATE suite (own encode vs. own decode of the exact same bits, so the higher SV8 number just reflects less inter-decoder disagreement, not different codec fidelity).
• ~1250× realtime (a 0.5 s clip decodes in ~0.4 ms; a 5.8 s clip in ~4.6 ms; Apple M-series, Node 25).

License

ॐ · BSD-3-Clause, matching the bundled libmpcdec.

Part of @audio/decode — the decode family umbrella.